Research On Rain Image Restoration And Object Detection Algorithm Based On Deep Learning

From the process of computer vision technology continues to pursue development,object detection is still full of arduous challenges in the present and future,including image quality degradation caused by media changes.If the image is acquired in rainy weather,the rain line will block and interfere with the object in the image,making the high frequency details and texture information disappear.Therefore,it has important application value to clear image restoration in rainy environment,and it is of great significance to improve the performance of object detection algorithm.On the basis of YOLOv5s object detection algorithm,this thesis optimizes the feature fusion stage and Mosaic data enhancement method,and uses the optimized U-Net rain image restoration algorithm as the front-end preprocessing to improve the detection performance of YOLOv5s detection algorithm in rainy environment.The specific work is as follows:（1）In terms of improving the performance of rain image restoration algorithm,this thesis uses multi-scale parallel cavity convolution to replace effective convolution,and enriches the extraction of rain line features on different sizes of receptive fields.At the same time,based on the idea of dense connection network,the feature fusion operation is carried out between the branches of each Res2Block module in the model to strengthen the flow of feature information between each branch and improve the utilization of features.In addition,to make the model distinguish feature values selectively,CBAM attention mechanism is added after network input,which improves the rain removal performance of the original U-net algorithm.（2）In terms of the loss function of the rain image restoration algorithm,this thesis introduces the L₁loss function and the structural similarity loss function which are more robust to outliers as punishments for the training of the rain network,and constrains the high-frequency details and texture information of the image restored by the network to have high similarity with the original clear image.The image restoration effect is improved to some extent.（3）In terms of improving the performance of the object detection model,this thesis uses the optimized U-Net rain image restoration algorithm as the preprocessing stage of the detection model,and on this basis,the Bi FPN feature fusion network is introduced as the feature fusion module of the model.At the same time,to make the trained model have better universality for the rainy environment,the Mosaic data enhancement method is optimized.In this thesis,rain image restoration experiments are carried out on synthetic raindrop dataset and real rain environment.It is proved that the optimized rain image restoration algorithm has better restoration effect than the original algorithm and other classical restoration algorithms.The image texture is perfect,and the structural similarity evaluation index is improved by 11.7%.At the same time,the detection experiment is carried out under the self-built rain object detection dataset,which proves that the optimized algorithm has better detection accuracy than the original YOLOv5s algorithm and other detection algorithms,and the object detection accuracy in the rain environment is improved by 9%compared with the original algorithm.